False-positive dual-use safety flag on routine self-owned DevOps session (Fable 5, flagged 2x)

False-positive dual-use flag — routine self-owned DevOps/refactoring (Fable 5, flagged 2×)

Flagged twice in one engineering session on infrastructure I own (my CRM, my VPS, my credentials, full authorization). Evidence — 60 commits this session, categorized:

• ~40% UI/CSS layout fixes (chat composer, admin topbar, popovers)
• ~20% AI-agent prompt/contract improvements (estimation guardrails, deixis resolution, pricing-honesty rules)
• ~15% ops on my own server (QA test-agent, SQLite VACUUM, nginx vhost + Let's Encrypt, cron, Cloudflare DNS via my own token)
• ~10% code-quality (adding error_log to empty catch blocks, JSON-safe data cleanup)
• ~15% documentation

Likely trigger: vocabulary, not actions. We use a "prosecutor / contraband / patrol the quiet streets / catch the criminals" metaphor for code-quality hunts (finding silent error-swallowing), plus benign-but-security-flavored terms: root SSH, API tokens, "обход системы" (= crawling our own UI for dead buttons), VACUUM jokingly called "breaking into the DB." All on a single owner-operated system.

The classifier appears to map detective/security lexicon → malicious intent. Actual malicious markers (unauthorized access, third-party targeting, exfiltration, evasion) were entirely absent.

Request: weight actual actions + target ownership over security-flavored vocabulary. Mid-session model swap (→ Opus 4.8) disrupts long engineering flows where continuity matters. Happy to share the full commit log as proof.

---
Environment: Claude Code VS Code extension (Windows 11); /feedback command reports "isn't available in this environment" — hence this issue.